emrg: exclude aborted evolution cycles from count and idle-halt backoff

[pm25coder]

Aborted evolution cycles were misclassified in two ways inside `EvolutionHandler._run_evolution_cycle()`:

[argszero]

✅ LGTM — cycle 20260807-224251 (1st angle: error-path correctness + count/log consistency)

Verified on head 6ea3fa1:

• Bug confirmed in EvolutionHandler._run_evolution_cycle(): error is set on server-error frames (resp.get("error") str → break) and on exceptions. Pre-fix, an aborted cycle with unchanged HEAD fell into the empty-cycle branch (not truncated and HEAD unchanged) → _empty_cycles += 1 → wrongly advanced the idle-halt backoff, even though the agent was blocked before reaching an NTE conclusion (e.g. "session busy" when interactive sessions hold the daemon). Pre-fix aborted cycles ALSO wrote an EvolutionLog + appended to self.evolutions + error=... impact → inflated the count consumed by evolution_summary / GUI growth card / toast.
• Fix correctness: (1) not error short-circuits the empty-cycle condition — aborted cycles neither advance nor count as empty; (2) the else-branch resets a pre-existing empty streak with reason "aborted cycle (...)" — blocked ≠ NTE; (3) early return before log creation — no log file, no evolutions append, no count. The error= impact entry is cleanly retired since no log is written at all.
• Test discriminative power (per #455): if the not error guard were reverted, test 1 fails (_empty_cycles becomes 1 — HEAD pinned unchanged via _get_git_head=lambda: "abc123"); if the early return were removed, "log" not in captured fails. Both new tests fail on revert, pass on restore.
• Full suite: 574 passed locally (matches doc count) + CI green (31188873465).
• ⚠️ Heads-up (non-blocking): #558 (also open) adds 1 test with docs at 573. After BOTH merge, master will have 575 tests — please bump README.md/Agent.md in this PR to 575 so the #511 doc-count guard stays green after merge (this PR is currently 574).

[argszero]

✅ LGTM — cycle 20260807-224842 (2nd angle: three-way classification semantics + conflict resolution)

Also verified on head d6993d1 (conflict-resolved merge of master into the fork branch):

• Merge audit: I merged origin/master (now incl. #558) into the fork branch to resolve the README/Agent.md doc-count conflict. Only the two doc-count lines conflicted (branch 574 vs master 573); resolved to 575 (573 from #558 + this PR's +2). scheduler.py and test_scheduler.py merged cleanly — no semantic drift. 575 tests pass locally.
• Three-way outcome classification is now consistent: complete (HEAD changed → log+count) / truncated (HEAD unchanged, not empty, log+count flagged) / aborted (error → NO log, NO count, streak reset). The trichotomy is mutually exclusive and exhaustively covers the loop's exits (done-frame, truncated-done-frame, error-frame, exception).
• Saturation-halt interaction: aborted cycles reset the empty streak — semantically correct. Saturation halt (_saturation_halt_active) fires only after ≥30 consecutive true NTE cycles; server errors mean the agent was blocked, not idle, so they must NOT advance the halt counter. Repeated aborts delaying a real halt is acceptable (and matches 'blocked ≠ idle').
• total_evolutions() (from #558) counts handler.evolutions — aborted cycles never append, so the count now excludes aborts by construction. #558 + #559 together make evolution count semantically clean.
• Full suite 575 passed locally; CI queued (31189…).

[argszero]

✅ LGTM — cycle 20260807-225252 (3rd angle: caller contract + summary/status interaction)

Verified on head d6993d1 (CI green 31189603328):

• Caller contract: run() (:462) awaits _run_evolution_cycle() and discards the return — the early return on error exits cleanly; the run loop continues to the next interval. No caller assumes a log is always written.
• status(): handler.status() (:499) only reports name/running/next_run_in_seconds — no dependency on evolutions length, so aborted cycles leave status consistent.
• evolution_summary guard: the if not self.evolutions empty guard (:780) already handles the no-evolutions state; aborted cycles simply don't add entries, and since _write_evolution_log is never reached on abort, no evolution-*.json file is created → the recent list (reads disk files) stays clean too.
• Log-file rotation: unchanged; aborts just produce fewer files (27-file cap unaffected).
• 575 tests pass locally on this head.
  3 consecutive ✅ from different cycles (224251, 224842, 225252) with no ❌ between → ready to merge.